Heterogeneous Graph Convolutional Networks And Matrix Completion For MicroRNA And Disease Association Prediction

MicroRNAs(miRNAs)are tiny RNAs that do not encode proteins and have a regulatory role in many cancers.Therefore,detecting the relationship between miRNAs and diseases is helpful for the diagnosis and treatment of diseases.In the past,researchers mostly used biological experiments to study the regulation of miRNAs on diseases,but the experiments were time-consuming and expensive.Using computational methods to predict miRNA-disease pairs with a high likelihood can reduce the number of experiments,thereby improving the accuracy and efficiency of biological experiments.Therefore,researchers began to study computational biology methods that use known miRNA-disease associations to predict potential miRNA-disease associations.The calculation methods for predicting miRNA-disease associations can be roughly divided into network-based methods,machine learning-based methods,and matrix-completionbased methods.The network-based method uses known miRNA-disease associations to build a topological graph,and finds the potential associations between miRNAs and diseases through the transfer of relationships between nodes in the topological graph,but new miRNAs or diseases are isolated nodes in the graph,This type of method cannot be used directly and needs to be improved.Machine learning-based methods use known miRNA-disease association training classifiers to predict new associations between miRNAs and diseases,but such methods need to design feature extraction methods and solve the problem of missing negative examples.The method based on matrix completion complements the missing values in the matrix by fitting the miRNA-disease association matrix,but ignores the information contained in the topological structure.The three types of methods have their own advantages and disadvantages.This article combines the three types of methods to propose a new calculation method that uses graph convolution and matrix completion on heterogeneous graphs to predict miRNAdisease associations.The method first fuses a variety of data to calculate the miRNA similarity matrix,uses the associations between the disease and the gene to calculate the disease similarity matrix,and then combines the two to construct a heterogeneous network through the known miRNA-disease associations.The graph convolution is used to extract the topological structure information on the heterogeneous network,and the expression of miRNAs and diseases in the latent variable space is obtained,so as to predict the relationship between potential miRNAs and diseases.We used four crossvalidation experiments to evaluate our method,and the verification results confirmed that our method is significantly better than the current method.In addition,we also did case analysis to further prove the effectiveness of our method.